DeepInfra launched its first international data center in Toronto on July 8, 2026, marking a significant expansion for the Palo Alto-based artificial intelligence firm. The facility introduces a high-density inference cluster utilizing NVIDIA Blackwell B300 hardware to the Canadian market, targeting enterprises that require localized compute resources.
The Toronto site provides 1.7 megawatts of capacity and houses more than 1,000 Blackwell B300 GPUs, specifically optimized for production-level inference workloads. This infrastructure allows DeepInfra to process real-time requests for over 200 open-source models, including Llama and Mistral variants, through its proprietary API stack. The company, founded in 2022 by Nikola Borisov, Georgios Papoutsis, and Yessenzhar Kanapin, maintains a vertically integrated model by owning its hardware rather than relying on hyperscale cloud providers.
Financial backing for this expansion follows a $107 million Series B funding round closed in May 2026. Investors in the round included 500 Global, Georges Harik, and NVIDIA, among others, signaling strong institutional support for the company’s focus on cost-efficient inference. The Toronto facility serves as the ninth global site for the firm and the first to operate outside the United States.
Canadian organizations in sectors such as finance and healthcare face rigorous regulatory frameworks, including the Personal Information Protection and Electronic Documents Act and Quebec’s Law 25. These regulations mandate strict controls over cross-border data transfers and processing. By establishing a physical presence in Toronto, DeepInfra aims to mitigate the latency and compliance friction associated with routing inference traffic to U.S.-based servers.
The Blackwell B300 hardware provides a distinct technical advantage for these inference-heavy tasks. With 288 gigabytes of HBM3e memory per unit, a single B300 can accommodate large models that previously required sharding across multiple H200 GPUs. This consolidation reduces communication overhead and hardware costs, which the company claims results in significant efficiency gains over general-purpose cloud alternatives.
DeepInfra’s infrastructure philosophy prioritizes the full stack, optimizing from the silicon level to the API layer. By managing its own GPU hardware, the company avoids the performance bottlenecks often associated with multi-tenant cloud environments. This approach is particularly relevant for the high-throughput requirements of modern generative AI applications, where token latency directly impacts user experience and operational costs.
The Toronto facility also benefits from the city’s mature colocation ecosystem, which allows for rapid deployment of high-density GPU capacity. While the company has not publicly identified the specific colocation provider, the site is designed to handle the thermal demands of liquid-cooled B300 units. This logistical readiness ensures that the cluster can scale to meet the growing demand for local inference capacity among Canadian enterprises.
Despite the physical residency of the servers, the legal status of the data remains subject to U.S. jurisdiction. The U.S. Clarifying Lawful Overseas Use of Data Act, or CLOUD Act, grants U.S. law enforcement the authority to compel domestic technology companies to produce information regardless of where the physical servers are located. Because DeepInfra is a U.S.-incorporated entity, it remains within the reach of these legal mandates, creating a distinction between data residency and true data sovereignty.
This legal reality presents a challenge for Canadian government entities and highly regulated industries seeking full autonomy over their data. Prime Minister Mark Carney recently emphasized the necessity of Canadian-controlled infrastructure to mitigate foreign strategic exposure, setting a target for 850 megawatts of sovereign compute capacity by 2030. DeepInfra’s Toronto cluster, while beneficial for latency and standard residency compliance, does not meet this threshold for sovereign control.
The distinction between residency and sovereignty will likely dictate future procurement strategies for sensitive Canadian workloads. While consumer-facing applications may find the Toronto cluster sufficient for their operational needs, organizations handling classified or highly sensitive information must weigh the benefits of local inference against the persistent legal reach of the CLOUD Act. The industry will monitor whether further regulatory shifts or the emergence of local providers will eventually bridge this gap in the Canadian compute market.
